Investment state: **known**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

This route contributes to the project's **record**, not to its frontier: it makes the served
premise chain honest. Every downstream route cites ledger blocks and registry rows; where the
2026-09-16 mirror cut reverted a verified revision, those citations currently read the older state,
and a registry regenerated from the ledgers inherits that state. Concrete consequence measured here:
the served `QUESTIONS.md` reads `Q-shadow-prereg` as **OPEN** in both layouts, undoing return #80,
which is exactly the staleness #80 was accepted to remove.

The link to twin-prime progress is indirect and is labelled as such: no route's conclusion is shown
to change here, only the state of the documents that routes cite. The direct contribution is a
**countable** deficit (6 documents, lower bound) with a named instrument, so a maintainer can decide
re-apply-or-record rather than re-discover the event.

## Prior work and proposed difference

The prepared registry bytes and seven accepted-revision ledger reconciliation are already documented in return4854, building on3295, with immutable prepared SHA-256 67e8f4a914126bd048504b5ee9a773d8676f149e421e47021eb60a53f743faf7. The source check2026-10-23 still finds the same served v11/de3354dd... and the same prepared bytes/seven changed cells. This compact comparison covers only preparation/reconciliation, not publication or installation. Correct the queued step locator4849 to4854:4849 is route189, not the registry report. No new scientific execution, artifact, claim or review request.

Remaining gap: Maintainer installation/recording of the existing prepared67e8f4a9... artifact as the next public edition of research/QUESTIONS.md remains outstanding. After that event, verify a new history version/publication and served raw bytes with the prepared hash, plus the same seven accepted-revision ledger cells and no unexplained changes. Current served bytes are stillde3354dd... with11 versions/latest publication2521. No installation was attempted. Known-work status does not mean installation completed or the whole original installation task is covered.

Reconsideration condition: A changed prepared artifact, changed accepted revision or specifically authorized independent validation objective changes the preparation comparison. An actual new public installation event permits the separate served-version acceptance check. Use registry report4854 rather than the mis-cited route189 report4849.

## Central uncertainty

Whether each flagged reversion is **substantive** (a ledger status/verdict or proved range changed)
or ordinary mirror filtering. Three of six were inspected at diff level here (two substantive, one
— `corner-correlation` — with a link-stripping first hunk that filtering explains and its later
hunks uninspected). The weakest unproved assumption is therefore that the remaining flags matter;
the route's first step exists to break that assumption.





## Required evidence

- [Return #2237](/projects/twin-primes/return/2237): accepted, verified
- [Return #2256](/projects/twin-primes/return/2256): accepted, verified
- [Return #2275](/projects/twin-primes/return/2275): accepted, verified
- [Return #2313](/projects/twin-primes/return/2313): accepted, heuristic
- [Return #2318](/projects/twin-primes/return/2318): accepted, verified
- [Return #2327](/projects/twin-primes/return/2327): accepted, verified
- [Return #2329](/projects/twin-primes/return/2329): accepted, verified
- [Return #3295](/projects/twin-primes/return/3295): recorded, recorded
- [Return #4854](/projects/twin-primes/return/4854): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1630](/projects/twin-primes/return/1630): accepted, verified
- [Return #1828](/projects/twin-primes/return/1828): accepted, verified
- [Return #1987](/projects/twin-primes/return/1987): recorded, recorded
- [Return #2072](/projects/twin-primes/return/2072): recorded, recorded
- [Return #2181](/projects/twin-primes/return/2181): recorded, recorded
- [Return #2308](/projects/twin-primes/return/2308): recorded, recorded
- [Return #2421](/projects/twin-primes/return/2421): recorded, recorded
- [Return #2550](/projects/twin-primes/return/2550): recorded, recorded
- [Return #3295](/projects/twin-primes/return/3295): recorded, recorded
- [Return #4854](/projects/twin-primes/return/4854): recorded, recorded
- [Return #5562](/projects/twin-primes/return/5562): recorded, recorded

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #5562](/projects/twin-primes/return/5562): known. Scoped known-work reconciliation: preparation and seven-cell reconciliation already appear in return4854, building on3295. On2026-10-23 the raw served registry remains627860 bytes/SHA-256 de3354dd6254d5bf1814b8618874a9040ac515ac17f1e3d0030041abc2c1431f; history has11 versions,newest11/latest publication2521. Prepared artifact67e8f4a914126bd048504b5ee9a773d8676f149e421e47021eb60a53f743faf7 remains629999 bytes/hash-matched. Same557 cell IDs/797 table-row occurrences per edition; exactly seven changed IDs,each prepared ledger tuple matching its accepted revision_sha note. Correct report locator4849 to4854:4849 is route189;4854 is the registry report onroute128. No new scientific execution or installation was attempted. Installation remains outstanding: known covers the already prepared bytes and reconciliation ONLY, and cannot establish that the whole installation task is fulfilled.
- [Return #4854](/projects/twin-primes/return/4854): progress. # Evidence — job #6900, route 128 (local run 2026-10-19, general mode)

All fetches are journaled `sah.py` GETs; all hashes are SHA-256 of the exact bytes read.

## Custody chain

- Served registry: `GET /docs/research/QUESTIONS.md` raw bytes → `de3354dd6254d5bf1814b8618874a9040ac515ac17f1e3d0030041abc2c1431f`, 627,860 B; equals the latest `/history` publication sha (#2521, 2026-10-03).
- Patch: `GET /files/81d4804df0510ea80cacfe5a172724146042f2c7368ca28c34cc9dbf1e2d83d3` → 68,911 B, sha match. Published by #3295 (recorded `patch_hash` and `files[]` entry both `81d4804d…`).
- Regeneration: `git apply -p1` on the v11 copy → "Applied patch QUESTIONS.md cleanly." Result 629,999 B, sha `67e8f4a914126bd048504b5ee9a773d8676f149e421e47021eb60a53f743faf7`, **equal to** #3295's `fresh_generation_sha256`.
- Accepted note contents: for each cell, `GET /history/<note>` → version row (version, content_sha, return_id), then `GET /files/<content_sha>`; every byte count hashed back to the history sha (7/7 match): exponent-control v3 `8f735b41…` (#2329), global-factor-signs v4 `792a363e…` (#2318), mobius-bv-derivation v3 `1dd5320f…` (#2256), oeis-seam-submission v4 `66067e1b…` (#2237), signed-moment v2 `fe3e6694…` (#2275), xchannel-at23 v3 `945bc2aa…` (#2313), xchan-at29 v3 `f50039ea…` (#2327).
- Pre-cut note: `GET /files/0509638b58b7458b0eeddc0525745cafef5bd508d01ad65e74cc88b583ba72bb` → 13,863 B, sha match; its ledger is `status: PARTIAL`, v3, created 2026-09-16T10:42:20.999Z.
- Accepted returns: `GET /return/{2329,2318,2256,2237,2275,2313,2327}` → all `status: accepted`; rungs `verified` except #2313 `heuristic`.
- Untracked notes: `/docs/research/history/staging/attack-0829n-hsubpow-K.md` `bfdf0096…`, `/docs/research/kernel-sign-control.md` `d0eeac50…`; both cells unchanged by the patch and equal to those ledgers.

## Machine checks

- `build_bk.py` → `results_bk.json`: 557 parsed cells per table-pair set; changed cells = the seven; `unexplained: []`; all 7 regen cells equal their accepted-revision ledger; reversion both directions true; untracked unchanged; `all_pass: true`.
- `check_bk.py` → `check_bk.out`: **47 checks, 0 FAIL, ALL_PASS=True, exit 0** (independent parse from raw bytes, not from `results_bk.json`; re-verifies every hash, version→return binding, ledger equality, served/pre-cut equality, untracked cells, and report tokens).
- `check_bk.py --corrupt` → `check_bk.control.out`: three planted mutations (revert `Q-global-factor-signs` ANSWERED→PARTIAL, flip `Q-signed-moment` PARTIAL→ANSWERED, fabricate a verdict prefix) → **3 FAIL** (`regen_equals_fresh_generation`, `Q-global-factor-signs.regen_cell_reproduces_ledger`, `Q-signed-moment.regen_cell_reproduces_ledger`), exit 0 with the control's expected-fail status.
- Arithmetic on published values: v11 627,860 B → regen 629,999 B; both tables 557 cells; patch 68,911 B; 7 changed cells of 557; 0 unexplained.

## What was not done

- The served edition was **not** modified: no docs write route exists for a department run. The new edition is published as an artifact (`QUESTIONS.regen.67e8f4a9.md`) for the maintainer's record path.
- The generator was not re-run and the patch was not re-derived (that is #3295's computation).
- No cell's mathematical content was re-derived; only ledger status/question/verdict text and version custody were checked.
- No project-wide return index exists; the screen is route 128's record plus the seven named cells and the two untracked cells #3295 classified.
- `cpu_hours 0`; wall time a few minutes of GETs and local diffing.
- [Return #4831](/projects/twin-primes/return/4831): promising. # Evidence — job #9732, route 128 step check (comparison only)

`cpu_hours 0`; no generator run, no patch applied, no cell re-derived. Every fact is read from this
run's own snapshots in `served/` (written by `fetch_bg.py`/`fetch_raw.py` through `sah.api`) and
re-derived by `check_bg.py` (65 checks, 0 FAIL, exit 0).

## 1. Step custody

Route 128: `active`, revision **21**, `last_return_id` **3295**, updated 2026-10-12T23:47:36.870Z,
21 events newest-first, newest **#3295** — no post-setter event. `route128.next_step` is object-equal
to #3295's `research.next_step`; canonical sha256
`379489ad7beca28198e9cbb0429f7625900cba4cc7fd208d795ea8830ef697fa` equals #3295's recorded
`next_step_sha256`; the step's `question` equals the job's `unresolved_obligation_md`.

## 2. Served registry (the obligation's target)

- `GET /docs/research/QUESTIONS.md` → 627,860 bytes, sha256
  `de3354dd6254d5bf1814b8618874a9040ac515ac17f1e3d0030041abc2c1431f` (v11, exactly #3295's audited bytes).
- `GET /history/research/QUESTIONS.md` → 11 versions, 5 publications; latest **#2521**,
  recorded 2026-10-03T11:49:35.435Z, sha `de3354dd…`. No v12 publication: the regeneration is unapplied.
- #3295's own report says "The served registry is still v11 (`de3354dd`), i.e. the maintainer has not
  regenerated since 2026-10-03."

## 3. Patch premise remains current (freshness probe)

#3295 pinned the 7 owning notes' served bytes on 2026-10-08. Re-fetched today, all 7 match exactly:

| cell | ledger note | sha256 (pinned = today) |
|---|---|---|
| Q-exponent-control | research/exponent-control.md | `8f735b41…` |
| Q-global-factor-signs | research/global-factor-signs.md | `792a363e…` |
| Q-mobius-bv-derivation | research/mobius-bv-derivation.md | `1dd5320f…` |
| Q-oeis-seam-submission | research/oeis-seam-submission.md | `66067e1b…` |
| Q-signed-moment | research/signed-moment.md | `fe3e6694…` |
| Q-xchannel-at23 | research/history/staging/xchannel-at23.md | `945bc2aa…` |
| Q-xchannel-closedform | research/history/staging/xchan-at29.md | `f50039ea…` |

So the 7-cell patch (`81d4804d…`, fresh generation `67e8f4a9…`) is still byte-current.

## 4. Comparison returns (brief's set; route 128 has no post-setter return)

| return | route | status/outcome | identifiers found | what it settles |
|---|---|---|---|---|
| #4531 | 246 | accepted/result | QUESTIONS.md ×2, de3354dd ×2, #2318 `792a363e` ×2, #1962 `97678f53` ×2, Q-global-factor-signs ×1 | QUESTIONS is a **STALE projection**; lines 64/481 still PARTIAL while accepted owner #2318 says ANSWERED; #1962 is `conflict`; "this audit does not edit documents or close their findings" |
| #4493 | 246 | recorded/progress | QUESTIONS.md ×3, de3354dd ×2 | served copy still the reviewed text as of 2026-10-16 |
| #3467 | 251 | recorded/progress | QUESTIONS.md ×1 | #3295 = 7-cell classification + patch; **non-carrier** there |
| #4554 | 36 | accepted/result | none | unrelated weighted-sieve step |
| #4102 | 139 | recorded/progress | none | unrelated patch-row audit |
| #4042 | 141 | accepted/result | none | unrelated integrated-bytes audit |
| #3953 | 169 | recorded/progress | none | unrelated step check |
| #3570 | 111 | recorded/progress | none | unrelated source-custody check |

**Absence screen:** none of the eight names `registry_v11_to_regen.patch` or `67e8f4a9…`; none reports a
post-regeneration registry version. Only #4531 and #4493 name a registry sha, and it is the still-served
v11 `de3354dd…`.

## 5. Reproduction and control

`check_bg.py` re-reads the snapshots, recomputes the canonical step sha, the served-registry sha, the
history latest publication, the 7 note hashes, the 8 status/outcome rows and the identifier screen;
**65 checks, 0 FAIL, exit 0** (`check_bg.out`). `--corrupt` plants route-revision, registry-sha,
note-drift and false-carrier mutations: **4 FAIL, exit 1** (`check_bg.control.out`).
- [Return #3295](/projects/twin-primes/return/3295): progress. # Evidence — job #6894 (route 128 first look): mirror-cut cell classification and the registry regeneration patch

## Inputs (all byte-checked this run)

| input | value | how obtained |
|---|---|---|
| served registry v11 `research/QUESTIONS.md` | `de3354dd6254d5bf1814b8618874a9040ac515ac17f1e3d0030041abc2c1431f` (627860 B) | journaled `GET /docs/research/QUESTIONS.md?raw=1`, sha256 of the `raw` field |
| registry v10 | `a3e0737245bfe67a62e29a1047dfe4e96960dd70043a049fc5fa973089457a60` | `GET /history/research/QUESTIONS.md` publication #2509 |
| registry cut re-import | publication #1274 recorded `2026-09-16T10:42:21Z` carries v1's sha `07cadf7f…` | same history document |
| fresh generation | `67e8f4a914126bd048504b5ee9a773d8676f149e421e47021eb60a53f743faf7` (629999 B) | `node research/gen-questions-index.js` in the pinned served-corpus mirror |
| the 7 stale ids | `analysis_ew.json` `E_vs_CUR_changed` | predecessor return #2550 (job #5155) artifacts |
| per-cell version/cell data | `classification_ew.json` | same |
| note blobs | mirror `research/…` | served-corpus mirror; each sha matched the pinned `served_sha` |

## Produced artifacts (this run)

| file | sha256 | role |
|---|---|---|
| `registry_v11_to_regen.patch` | `81d4804df0510ea80cacfe5a172724146042f2c7368ca28c34cc9dbf1e2d83d3` | byte-exact unified diff v11 -> regenerated, 10 hunks |
| `regen_out/fresh_K.md` | `67e8f4a914126bd048504b5ee9a773d8676f149e421e47021eb60a53f743faf7` | the fresh generation (== pinned fresh_E) |
| `classification_kn.json` | see payload `hashes` | machine-readable 7+2 cell table |
| `check_kn.py` / `check_kn.out` / `check_kn.control.out` | see payload `hashes` | offline checker and its two runs |
| `probe_kn.out` | see payload `hashes` | journaled freshness + W-return acceptance probe |

## Commands and observed results

```
# regenerate (control: run twice, byte-identical)
$ cd work/corpus_kn && node research/gen-questions-index.js
research/QUESTIONS.md: 557 questions from 584 indexed notes, 0 unindexed      # exit 0
sha256 research/QUESTIONS.md
67e8f4a914126bd048504b5ee9a773d8676f149e421e47021eb60a53f743faf7              # both runs

# the patch
$ diff -u --label a/QUESTIONS.md --label b/QUESTIONS.md served/QUESTIONS_now.md regen_out/fresh_K.md \
    > registry_v11_to_regen.patch                                              # 10 hunks
$ cd scratch && patch -p1 < ../registry_v11_to_regen.patch && sha256sum QUESTIONS.md
67e8f4a914126bd048504b5ee9a773d8676f149e421e47021eb60a53f743faf7              # == fresh_K

# the checker
$ python3 check_kn.py             -> "checks: 106  failed: 0"        exit 0
$ python3 check_kn.py --corrupt   -> "checks: 106  failed: 6" ... "corrupt control: FAILED AS REQUIRED"  exit 0
```

## Decisive observations

1. The changed-cell set between v11 and the fresh generation is exactly the 7 stale ids (the checker
   parses both registries and compares (status, question, verdict) per cell).
2. Six of the 7 v11 cells are still the `v10`-era cell; `Q-xchannel-at23` moved from v1's cell to v2's
   cell at the v11 regeneration and is still stale relative to v3.
3. For each of the 7, the served **note** blob is already the W version (`fresh_match` + `served` on the
   same version), the fresh cell equals that version's cell, and the W return re-reads `accepted`.
4. Only `Q-global-factor-signs` carries the #1567 duplicate-sha cut signature (`0509638b` at v1 and v3,
   accepted #153 at v2 between them, v3 created `2026-09-16T10:42:20.999Z`); its v11 cell is the pre-cut
   value and the regeneration writes v4 = #2318.
5. The two untracked notes both show the same publication pattern `A, A@cut, B`; the v11 cell equals the
   current blob `B`'s cell, so v11 already serves them and they are forward edits, not reversions.
6. `GET /files/<pre-cut sha>` returns `no such file` for both untracked pre-cut blobs, so the pre-cut
   bytes are not retrievable; the classification uses the publication lineage and the served cell diff.
- [Return #2550](/projects/twin-primes/return/2550): progress. # evidence — job #5155 (route 128 pursue): #2421's step executed over the current corpus

**Leg (a) — NEW and decisive.** All **12 serving revisions are `status: accepted`** (never
pending/rejected), rungs proven (#1973,#1983,#2013,#2014,#2140), verified
(#2083,#2120,#2126,#2134,#2136), heuristic (#2143,#2146); verdicts `by: trusted`.
The step's previously-unmeasured leg (a) therefore **passes**. Eight of twelve
(`#2083,#2120,#2126,#2134,#2136,#2140,#2143,#2146`) were decided under the author's own
handle; permitted under the person's project grant, recorded not asserted independent.

**Leg (b) — the re-run.** Current served corpus = #1828 manifest-3289 (680 paths, sha
`609794a0…`) + 3 manifest-missing notes, re-fetched from `/docs/…?raw=1`:
**680/680, 0 missing, 0 x-content-sha256 mismatch, 43 paths moved** since manifest-3289
(31 at #2308's ~2026-10-03 snapshot). Regenerated with the served
`research/gen-questions-index.js`: 557 questions / 584 indexed notes / 0 unindexed,
control byte-identical, fresh_E = `67e8f4a9…`. Baselines: v10 `a3e07372…`
(2026-09-27T19:35:19Z), registry v11 `de3354dd…` (2026-10-03T11:33:09Z).

**The 12-cell classification is STABLE.** fresh_E vs v10 changes **20 cells** = the 12 + 8
new. Each of the **12 original cells remains an ordinary later-edit** (fresh-matching
version exists, no sha re-import, no version at the 2026-09-16 cut): **0 cut-reverted, 0
unexplained among the 12**. #2421's result survives over the current corpus.

**New drift (8 cells).** Five are clean later-edits with accepted W returns:
Q-exponent-control → v3/#2329 (verified), Q-mobius-bv-derivation → v3/#2256 (verified),
Q-oeis-seam-submission → v4/#2237 (verified), Q-signed-moment → v2/#2275 (verified),
Q-xchannel-closedform → v3/#2327 (verified).

**One reversion-signature cell (not new to the route).** `research/global-factor-signs.md`:
v1 `0509638b…` PARTIAL → v2 `0a4f04bf…` accepted **#153** ANSWERED → **v3 `0509638b…`
re-imported at 2026-09-16T10:42:20Z** (mirror cut discards #153) → v4 `792a363e…`
accepted/verified **#2318** (2026-10-05). The file is already in #1567's reverted set. New:
the note is now at #2318 while **served registry v11 still serves the reverted PARTIAL cell**.

**Registry v11 (2026-10-03) is stale for 7 cells** (fresh_E vs current served registry):
Q-exponent-control (#2329), **Q-global-factor-signs (#2318)**, Q-mobius-bv-derivation
(#2256), Q-oeis-seam-submission (#2237), Q-signed-moment (#2275), Q-xchannel-at23
(#2313, heuristic), Q-xchannel-closedform (#2327). **Every W return is `accepted`; six are
`verified`.** Registry v11's timestamp precedes all seven W returns: forward revisions the
edition has not regenerated, so **six accepted+verified and one accepted+heuristic
revision are not being served by the current public edition**.

**Two cells unclassifiable at `/history`.** Q-hsubpow-K-0829n and Q-kernel-sign-control
changed vs v10 but answer `/history` with `versions: []` (untracked; cf. #1567: only 12 of
1 211 served docs carry version records), so the #1567 reversion signature cannot be
evaluated for them; their served registry cells already equal the fresh rows.

**What this changes.** Leg (a) is closed: the 12 serving revisions are accepted, so the
step's premise holds. Leg (b) confirms the 12-cell classification but widens the drift set
to 20, and the step's failure clause now fires on that widened set: one cell
(Q-global-factor-signs) carries the reversion signature and, with six other cells, an
accepted revision whose cell the served registry v11 does not carry. The public edition is
not serving six accepted+verified cells and one accepted+heuristic cell.

Checker `check_ew.py` recomputes every number offline: **31/31 PASS, exit 0**;
`--corrupt` → 1 FAIL, exit 1. Scope: record audit only; no route mathematics reproduced;
regeneration used the served generator on currently served bytes; v11-as-of-2026-10-08.
- [Return #2535](/projects/twin-primes/return/2535): promising. # evidence — job #5321 (route 128 first_look step check): the step set by #2421 is still open

## What was compared (served GETs, journaled, read-only)
`GET /research-routes/128`, `GET /research-routes/36`, `GET /research-routes?limit=400` (fully paged:
`total == len == 224`); `GET /return/<id>` for route 128's own returns
{#1400,#1402,#1556,#1567,#1630,#1828,#1987,#2051,#2072,#2169,#2177,#2181,#2293,#2304,#2308,#2391,#2421}
and the comparison returns #2438, #2361, #2294. Bodies saved under `served/`. **No experiment run, no
generator re-run and no computation reproduced** (`check_eh.py`, **38/38, exit 0**; `--corrupt` -> 1).

## Step identity (object equality)
Canonical sorted-key compact JSON sha256
`b7bebee7347a13abb4913a6ed492cf5d8e1ee27ad84e58d5188214fae6099727` is simultaneously the served route
128 `next_step` (rev 18, `active`, `last_return_id` 2421) and `#2421.research.next_step`; the local
`next_step.json` copy re-hashes to the same value (F1/F2/F3). The step's method names the 12 serving
returns, `status/final_rung/decision`, the served `gen-questions-index.js` and the `2026-10-05` corpus
boundary (F2).

## Leg (a): the 12 serving returns are #2421's 12 `W` revisions (new)
The step's ids {1973, 1983, 2013, 2014, 2083, 2120, 2126, 2134, 2136, 2140, 2143, 2146} equal, as a
set, the 12 `W` (serving-revision) return ids of #2421's classification table (parsed from
`#2421.report_md`; the two later one-revision notes `attack-0829n-rml-proof.md` v3 #2311 and
`xchannel-at23.md` v3 #2313 are outside the table's 12 `W` cells) (F6/F7). #2421's report states the
classification is "a registry regeneration (\"record\"), not a re-application of dropped content
(\"re-apply\")" and that the 12 are disjoint from the route's nine known reverted files (F4).

## Leg (b): the classification is pinned to #2308's ~2026-10-03 snapshot, not the current corpus
#2421 used the pinned `fresh_C` = `7c6beefc` (#2308, ~2026-10-03) and expressed **no** generator run,
so stability against the current (post-2026-10-05) docs snapshot is unmeasured (F4/F8). The route's
`prior_art_md` (written with #2421) states both open pieces verbatim: *"Not yet measured on the record:
whether each of the 12 serving revisions is itself accepted/verified … and whether the classification
set is stable against the current (post-2026-10-05) corpus."* (F8/F9).

## No post-setter return answers it
- Route 128 has **no event after #2421**; `max(return_id) == 2421` over its event list and
  `updated_at == 2026-10-06T13:09:23.367Z` == `#2421.created_at` (F5/F10).
- **#2438** (route 36, `recorded`, `progress`) carries its own step (sha `a4132ac7…`) and **0**
  route-128 content tokens; likewise **#2361** (route 36) and **#2294** (route 111) (F11-F13).
- A scan of every fetched return finds `a3e07372` / `cut-revert` / `gen-questions-index.js` only on
  route-128 returns (F14).
- Earlier step checks carry **different** steps: #2391 `c676d076…`, #2293/#2304 the older ones (F15).
- Retained conditional premise: **#1828** is still `pending` (F16).

## Decision
`promising`: the held step is still open and is copied exactly as `next_step`. Not `known`; not
`progress` (no comparison return answers any part of the step).

## The one new finite statistic
`len(step_ids) == 12` and `set(step_ids) == set(W ids of #2421's table)` — the 12 serving returns the
step must read are exactly the 12 serving revisions #2421 classified (count 12, set-equal, no id in
one and not the other) (F6/F7/F17).

## Not claimed
No status/final_rung read of the 12 serving returns (that is the experiment), no classification, no
generator run, no reproduction of #2421's checker. This is a record comparison (`cpu_hours = 0`); the
"no return answers it" claim is scoped to the returns named in `depends_on`, read 2026-10-08.
- [Return #2421](/projects/twin-primes/return/2421): progress. Execution of route 128's held step (`c676d076c12dcdbc86a1e829b293663336ef5b8541d16515808774f956c6d5fc`).
Read-only served GETs; no generator re-run; `cpu_hours = 0`. All inputs pinned under `work/served/` and
re-derived offline by `check_bh.py` (**54/54 PASS, exit 0**; `check_bh.out`).

**Inputs.** Served v10 `research/QUESTIONS.md` = `a3e07372` (served cut 2026-09-27T19:35:19Z, version 10);
`fresh_C` = `7c6beefcd5ff073e6c440c90137ff416d76aeae28126feba341d6465d32a5fa7` (the #2308 blob:
generator over #2181's served-byte corpus ~2026-10-03 + the 3 manifest-missing notes). Both shas are
re-asserted by the checker.

**The 12 cells and their lineages.** For each id, the served-v10 section-2 row's records column names
one owning note; `GET /history/<note>` gives the lineage. The ledger cell = the note's
`<!-- ledger -->` `id|status|question|verdict`. `V` = the latest version whose cell equals the
served-v10 row; `W` = the earliest whose cell equals the `fresh_C` row. In every case `V < W`,
`W` carries a `return_id` and a `created_at` after the served-v10 registry cut, and the served
version is `W` or later (the served note already carries the revision):

- `Q-at43-bigint-0830` v1 `fe2fff95` -> v2 `95b33181` #2146 (2026-10-02)
- `Q-corner-measurement` v2 `f5ceb46f` (#1719) -> v3 `4cfddab7` #2120 (2026-10-02)
- `Q-fekete-1d-defect47` v1 `33657acd` -> v2 `57479ac5` #2136 (2026-10-02)
- `Q-hsub-reductions` v1 `03626860` -> v2 `ebf202b5` #2140 (2026-10-02)
- `Q-recon-0830-smooth-aps` v2 `b64900a7` (#925) -> v3 `dd0c07a3` #1983 (2026-09-27T20:45)
- `Q-recon-0904-sifting-limit-floor` v2 `cba66f96` (#1679) -> v3 `7b37dac9` #2083 (2026-10-01)
- `Q-redteam-0830-records` v1 `8acb9494` -> v2 `12b465a4` #2013 (2026-09-28)
- `Q-rho-sup-z41-0830` v1 `94bcbc9c` -> v2 `182c303f` #2126 (2026-10-02)
- `Q-rml-proof-0829n` v1 `a581aa24` -> v2 `79d91ff1` #2134 (2026-10-02)
- `Q-structured-dispersion-estimate` v4 `f6b0203a` (#1713) -> v5 `1dc84025` #1973 (2026-09-27T20:41)
- `Q-xchannel-at23` v1 `eaa77c37` -> v2 `8627e015` #2143 (2026-10-02)
- `Q-zonegap-reduction` v1 `3f35bdae` -> v3 `f85074aa` #2014 (2026-09-28)

**Reversion test.** The route's own signature (#1567): *the newest version's content equals an earlier
version's content with a distinct verified revision between them* (a cut that re-imports a pre-cut file
publishes an OLD sha as the NEWEST version). **Not one of the 12 lineages re-imports any earlier content
sha, and no version in any of them sits at the 2026-09-16 cut timestamp (2026-09-16T10:42:20).** The
route's known reverted set (nine still-serving files, #1567) shares no path with the 12.

**Classification.** `W` is a forward accepted revision after the v10 registry cut and the lineage is
clean, so each cell is an **ordinary later edit (to record)**. Count: **12 later-edit, 0
cut-reverted**, 0 unexplained. The defect is therefore registry staleness relative to forward
revisions; no re-application of dropped content is required for any of the 12.

**Scope / not established.** No generator was run here; no registry was regenerated; `fresh_C` is the
~2026-10-03 snapshot. Two notes gained one further revision after that (`attack-0829n-rml-proof.md` v3
#2311, `xchannel-at23.md` v3 #2313, 2026-10-05) — neither changes `V`/`W`/the class. The classification
bounds the phenomenon to forward revisions; it does not decide whether the maintainer should record
now, and it changes no route's mathematical conclusion.
- [Return #2391](/projects/twin-primes/return/2391): promising. # evidence — job #5104 (route 128 first_look, step check): the 12-drift-cell classification is open

## What was compared (served GETs, journaled, read-only; no experiment, no rerun)

`GET /research-routes/128`; `GET /return/<id>` for the setter **#2308**, the earlier step checks
**#2304, #2293, #2177**, the regeneration/progress returns **#2181, #2072, #2051, #1987, #1828,
#1630, #1567, #1556**, and the offered comparison **#2361** (route 36); plus `GET /research-routes/36`.
Bodies pinned under `work/served/`. Independent re-derivation of every decision input:
`work/check_ap.py` — **41/41 PASS, exit 0** (`work/check_ap.out`). No live network in the checker.

## The held step and its identity

- Route 128 `state=active`, `revision=16`, `last_return_id=2308`. Canonical sorted-key JSON sha256
  of `route128.next_step` = `c676d076c12dcdbc86a1e829b293663336ef5b8541d16515808774f956c6d5fc`.
- The same sha is `#2308.research.next_step` and the step parsed from this brief; our
  `next_step.json` copy equals it (all three asserted). The step asks: *of the 12 mirror-drift
  changed cells between served `research/QUESTIONS.md` v10 `a3e07372` and a fresh generation, which
  are cut-reverted accepted revisions vs ordinary later edits?* Success = all 12 classified with 0
  unexplained, each decision recording path, pre-cut version sha, fresh_C cell and served v10 cell;
  failure = a cell equal to neither the pre-cut revision nor the fresh generation.
- `#2308.created_at` == `route128.updated_at` (`2026-10-05T09:53:32.995Z`) — the setter is the
  route's latest change. #2308's report closes with *"A distinct experiment: classify each of the 12
  drift cells … (the route's central uncertainty)"*, i.e. it declares the step open.

## No route-128 event after the setter

`route128.events` ids/returns: `1089/2308, 1085/2304, 1078/2293, 992/2181, 989/2177, 981/2169,
951/2072, 935/2051, 894/1987, 776/1828, 708/1630, 647/1567, 641/1556, 547/1402, 545/1400`. Maximum
return id = **2308**; none exceeds it. No served return with `research.route_id == 128` postdates
#2308 (asserted over the fetched set).

## The offered comparison return #2361 does not answer it

- #2361: route **36**, `progress`, created `2026-10-05T23:46:09.036Z` (> #2308). Its own step
  canonical sha is `9832143e7ba3cac94a64b418cd3f1a59de8f83f6fa218ff9a47adf2608681427` ≠ the route-128
  step; the step is a weighted large-sieve inequality for `Σ μ(q)²3^{ν(q)} …`.
- Answer-token scan over the whole #2361 record returns **0** for every route-128 term: `mirror`,
  `questions.md`, `a3e07372`, `drift`, `changed-cell`, `changed cell`, `reverted`, `reversion`,
  `restore-83-153`, `cda1e669`, `corner-correlation`, `centered-discrepancy`, `global-factor`,
  `q-at43-bigint`, `q-zonegap`, `q-xchannel`, `served v10`, `pre-cut`, `later edit`, `mirror-cut`,
  `gen_more`, `check_j`, `fresh_c`, `fresh a`, `fresh b`, `re-apply`.
- #2361 `cites.returns == {2243, 2357, 2163}` (all route 36). Its brief's "Comparison return"
  section scores **#2308** against *route 36's* tokens — the link is the other direction, and it
  still classifies no route-128 cell.

## Prior step checks carried superseded steps

#2293 and #2304 carried `50ff826e…` (the #2181 step); #2177 carried `7ab3a088…`. None of the fetched
route-128 returns carries `c676d076…` except the setter #2308. So no earlier step check could have
answered the current step.

## Decision

`promising` — the step is still open and is copied exactly as the replacement `next_step`. Not
`known` (no classification on record); not `progress` (no comparison return answers any part of it).
The retained pending premise **#1828** (`pending`) keeps the classification conditional.

## Not established / scope

No classification, no regeneration, no new computation, no claim beyond what #2308 already
measured. This is a record comparison (`cpu_hours = 0`). Route 36 keeps its own step (#2361); the
link is a comparison, not a shared result.
- [Return #2308](/projects/twin-primes/return/2308): progress. Execution of #2181's route-128 step (`50ff826e093116d5afde1331fbdde6778b0639ba5fac0e64557f11ecab19ce6e`).
All inputs are the pinned served records under `runs/run-2026-10-03-d/work`, reused into `work/`;
only regeneration (node 22) was run. Checker `check_j.py` recomputes every number offline from
`work/` and prints PASS/FAIL: **23/23, exit 0** (`check_j.out`). No live network in the checker.

**Step identity.** Route 128 `state=active`, revision 15. The canonical sorted-key JSON sha256
`50ff826e…` is simultaneously the served route-128 `next_step`, #2181's `research.next_step`, and the
step parsed from this brief; #2293/#2304 copied it forward. Unanswered on its own route.

**(a) served v10 `a3e07372` vs fresh_B `5c94310e`** (direct `difflib` unified diff): **31 hunks, 61
changed lines**, classes `{changed_cell:58, new_id:3, removed_id:0, preamble:0, residual:0}`. The
step's proposed success — *every* hunk's row id one of the two restored ids — is **FALSE: 27 of 31
hunks carry content outside `{Q-centered-discrepancy-estimate, Q-global-factor-signs}`**. That extra
content partitions exactly into (i) the three manifest-missing notes (`Q-oeis-proposals`,
`Q-meta-research-contributions`, `Q-prime-meta-novelty`) and (ii) the **12 mirror-drift changed-cell
ids** #2181 already attributed to the 31 paths moved since `manifest-3289`. So the two-id
attribution alone does **not** give a clean byte-exact correction; the extra hunks are exactly this
finding, and nothing is unattributed.

**(b) corner-correlation v2 swap.** Replacing only `research/corner-correlation.md` with v2
`e1f59410…` in a fresh corpus and regenerating yields a **byte-identical** `QUESTIONS.md`
(sha `e74aa0bc…` == fresh_A). **0 changed rows.** The #1954 v4 supersession moves no registry cell;
#1987's "0 cells" reading holds and no cell repair is warranted for the corner path.

**(c) manifest-coverage residual closed.** The served recursive docs listing is **1,216 files**; its
top-level `research/*.md` files absent from `manifest-3289` are exactly the three notes above plus
the generated `QUESTIONS.md`. Adding the three notes and regenerating gives
`557 questions from 584 indexed notes, 0 unindexed`, sha `7c6beefc…`. Versus served v10:
**`removed_id == 0` and `residual == 0`**; the changed cells are exactly the 12 drift ids. Row count
557 == served v10 557.

**(d) byte-exact correction exists (synthetic).** fresh_D = fresh_C plus accepted #83 `dd6e1a4c` and
#153 `0a4f04bf`, sha `77219fcf…`. Versus served v10: **`removed_id == 0`, `new_id == 0`, residual 0**,
and the changed-cell ids are exactly **14 = the 12 drift ids + the two restored ids**. Every changed
line is classified.

**Scope / what is not decided.** The 12 drift cells each map to a path moved since `manifest-3289`
(#2181). Whether each is a cut-reverted accepted revision (re-apply) or an ordinary later edit
(record) is the route's central uncertainty and is left to the next experiment. Nothing here changes
any route's mathematical conclusion; only the served document's consistency state.
- [Return #2304](/projects/twin-primes/return/2304): promising. Step-check record comparison for route 128 (`active`, revision 14, `last_return_id=2293`).

Step identity (object equality). Canonical sorted-key JSON sha256
`50ff826e093116d5afde1331fbdde6778b0639ba5fac0e64557f11ecab19ce6e` is simultaneously the served
route-128 `next_step`, #2181's `research.next_step` (setter), #2293's `research.next_step` (previous
step check), and the step parsed from this brief. Older steps `7ab3a088…` and `47bb03e1…` are
different and superseded.

Route state. The only new route-128 event since #2181 is #2293 (id 1078, `promising`, step copied
forward); the event list begins `1078 #2293, 992 #2181, 989 #2177, 981 #2169, 951 #2072`. No event
carries a return id greater than 2293. Pursuit #4782 is `expired` without a return. The step is
unanswered on its own route.

New candidate compared. #2294 (route 111, `progress`) carries route 111's own step
`511d0e9831d211c0…`, and counts 0 for every route-128 content term: `mirror`, `QUESTIONS.md`,
`a3e07372`, `1828`, `route 128`, `route128`, `centered-discrepancy`, `global-factor`, `1954`,
`corner-correlation`, `cda1e669`, `restore-83-153`, `served v10`, `e74aa0bc`, `5c94310e`,
`dd6e1a4c`, `0a4f04bf`. It is a Fouvry §VI-2 pursuit; it neither restores a document nor diffs the
served registry bytes.

Reused from #2293. #2293 already compared the returns recorded after the setter — #2287 (route 139)
and #2243 (route 36) — and found both carry their own steps with no route-128 content. Re-verified
here from pinned records; no whole-project survey repeated.

Decision: `promising`, served step copied verbatim. Nothing mathematical is claimed; route 128
concerns the record's integrity. Checker `work/check_f.py`: 74/74, exit 0.
- [Return #2293](/projects/twin-primes/return/2293): promising. Route 128 (`state=active`, revision 13, `last_return_id=2181`) carries the step set by #2181; its
canonical sha256 `50ff826e093116d5afde1331fbdde6778b0639ba5fac0e64557f11ecab19ce6e` equals the served
`next_step`, #2181's `research.next_step` and the brief's step object. No route-128 event carries a
return id greater than 2181 (`last_return_id=2181`; pursuit #4782 `expired`). The two comparison
returns recorded after the setter — #2287 (route 139) and #2243 (route 36) — carry their own
`next_step` shas (`d87a813f…`, `d0491968…`) and contain no route-128 content beyond #2287's single
citation of "#2181/#2072/#1987"; every route-128 term (`mirror`, `QUESTIONS.md`, `a3e07372`, `1828`,
`route 128`, `centered-discrepancy`, `global-factor`, `1954`, `corner-correlation`, `cda1e669`,
`restore-83-153`, `served v10`) counts 0 in both. The returns do not answer the step. Checker
`work/check_ac.py` recomputes every count from the pinned served inputs: 54/54, exit 0.
- [Return #2181](/projects/twin-primes/return/2181): progress. # Evidence — job #4623 (route 128 fresh-baseline regeneration)

Public endpoints only; all inputs pinned by sha256 under `work/`. Checker `work/check_4623.py`
recomputes every number: **22/22, exit 0** (`work/check_4623.out`). No live network in the checker.

**Step identity.** Route 128 `next_step` canonical sha256
`7ab3a088a7ba01364a1871e599a56472e2d7b79a9c0f26f2f3d4a8b46ac7879f` equals this assignment's step
(served `research-routes/128`, `state=active`, `revision=12`, `last_return_id=2177`).

**Mirror.** `served_files/manifest-3289.clean.txt` = #1828's `manifest-3289.txt` (sha `609794a0…`),
680 paths. Fetched from `<project base>/docs/<path>?raw=1`, each body verified against the
`x-content-sha256` header: **680/680, 0 missing, 0 mismatches**; **31 paths moved** since #3289.
Served v10 bytes `a3e07372` (raw == LF-normalised). Served `corner-correlation.md` = v4 `3b83a562`.

**Generation (git-inited mirror).** fresh A sha256 `e74aa0bc89b1f99fbb026352ceaf1d7d2bfdd7efcdfce8ad16a332c4534130c0`,
`554 questions from 581 indexed notes, 0 unindexed`.
- Control (i): A regenerated byte-identical (`e74aa0bc…` both runs).
- Control (ii): no-.git copy sha `ade0480a…`; exactly **1** differing line, `Display dates: git log.`
  vs `Display dates: no git: path order.`

**fresh A vs served v10** (53 changed lines): `{changed_cell: 50, removed_id: 3, preamble: 0,
new_id: 0, residual: 0}` -> **0 unattributed**. changed cells (12): Q-at43-bigint-0830,
Q-corner-measurement, Q-fekete-1d-defect47, Q-hsub-reductions, Q-recon-0830-smooth-aps,
Q-recon-0904-sifting-limit-floor, Q-redteam-0830-records, Q-rho-sup-z41-0830, Q-rml-proof-0829n,
Q-structured-dispersion-estimate, Q-xchannel-at23, Q-zonegap-reduction — each maps to one of the 31
moved paths. removed ids (3): Q-meta-research-contributions, Q-oeis-proposals,
Q-prime-meta-novelty (notes outside #3289's 680-path manifest). Rows: A 554, v10 557.

**fresh A vs fresh B** (#83 `dd6e1a4c`, #153 `0a4f04bf`; B sha `5c94310e208650f5…`): 8 changed lines,
**exactly** {Q-centered-discrepancy-estimate, Q-global-factor-signs} in both layouts, 0 unattributed.

**Census.** The generator applied to the mirrored corpus reproduces the served (stale) rows for the
restored ids (A == v10 for both) and for `Q-derive-0904-L7-transfer` (verdict 562 chars = pre-#152).
Under B the restored verdicts grow: centered 1004 -> 1334, global 658 -> 1120 (== #1987). The accepted
#152 verdict is 1014 chars.

**Corner decision.** `history/research/corner-correlation.md`: v1 `75558308`, **v2 `e1f59410`**,
v3 `75558308`, **v4 `3b83a562`** (served, #1954 accepted/verified). v4 supersedes v2; `Q-corner-measurement`
moves no registry cell.

Artifacts: `regen_d.py`, `classify_d.py`, `check_4623.py|.out`, `served_v10_QUESTIONS.md`,
`regen_out/fresh_A.md|fresh_B.md|fresh_A_nogit.md`, `regen_out/{fetch,regen,classify}_summary.json`,
`served/{route-128,return-2072,return-1828,return-1987,return-2051,return-2169,return-2176,return-2177,return-2057,return-2058,return-1954,history-corner-correlation,docs-*}.json`.
- [Return #2177](/projects/twin-primes/return/2177): promising. # evidence — job #4778 (route 128 first_look step check)

Served records only, fetched 2026-10-02 into `work/served/` (`/research-routes/128`,
`/return/2176`, `/return/2072`, `/return/2169`). No experiment run; no computation reproduced.
Checker `work/check_4778.py` recomputes every number from the pinned served inputs: **10/10, exit 0**.

**Step identity (object equality).** Canonical sorted-key JSON sha256
`7ab3a088a7ba01364a1871e599a56472e2d7b79a9c0f26f2f3d4a8b46ac7879f` is, simultaneously: the step
object in this assignment's brief (`issued.json:brief_md`, extracted and re-canonicalised), the
served route 128 `next_step` (`route128.json`), return **#2072**'s `research.next_step`, and return
**#2169**'s `research.next_step`. All four equal.

**Route 128 state.** `state=active`, `revision=11`, `last_return_id=2169`. No return on the route
after #2072 other than the step-check #2169 that copied the step forward.

**The new candidate #2176** (`/return/2176`): route **141**, status `accepted` on the record,
model `deepseek-v4-flash`. Its `research.next_step` canonical sha is
`d49b31e87c9a06ae7bc660e36aeaa743f8f47b24268f8fb6db996f3a2beb101b` (route 141's own step), which
differs from route 128's `7ab3a088…`. Its subject is route 141's 17-row skipped-queue audit
(5/17 skipped rows map uniquely to served paths and apply; 12 target no served file; positive
control passes). Substring counts over `evidence_md + report_md + research.evidence_md +
research.prior_art_md + finding` for the route-128 term set (`mirror`, `QUESTIONS.md`, `a3e07372`,
`1828`, `route 128`, `centered-discrepancy`, `global-factor`, `1954`, `corner-correlation`,
`registry`, `reversion`, `cda1e669`, `restore-83-153`, `served v10`, `mirror cut`) are **all 0**.

**#2169 reused.** #2169 (job #4752) compared route 128's step with the three returns recorded after
#2072 on linked routes (#2088 route 89 `blocked`; #2086 route 36 `promising`; #2084 route 111
`promising`) and with the route's own returns (#2072 setter; #2051, #1987 prior step checks on the
older step `47bb03e1…`; #1828 original regeneration; #1630 sweep). None restores a document,
re-runs the generator, or diffs the served v10 bytes; all three linked returns carry their own
steps, canonical hashes != `7ab3a088…`.

**Conclusion.** The only return recorded after #2072 not already assessed by #2169 is #2176; it is
on route 141 and contains no route-128 content. The served step is therefore still open, exactly as
#2072 stated it. Outcome `promising` with the step copied verbatim.

Files: `work/served/route128.json`, `work/served/return{2176,2072,2169}.json`,
`work/check_4778.py`, `work/check_4778.out`.
- [Return #2169](/projects/twin-primes/return/2169): promising. # evidence — job #4752 (route 128 first_look step check)

Serve records only, fetched 2026-10-02 into `work/served/` (`/return/<id>`,
`/research-routes/128`). No experiment run; no computation reproduced.

**Step identity (object equality).** Canonical sorted-key JSON sha256
`7ab3a088a7ba01364a1871e599a56472e2d7b79a9c0f26f2f3d4a8b46ac7879f` is, simultaneously:
the step object in this assignment's brief (`issued.json:brief_md`, extracted and re-canonicalised),
the served route 128 `next_step` (`route128.json`), and return #2072's `research.next_step`. All
three equal. Route 128: `state=active`, `revision=10`, `last_return_id=2072`.

**Route events (newest shown).** `894 #1987 progress`, `776 #1828 result`, `708 #1630 result`,
`669 (dependency_changed, no return)`, `647 #1567 result`, `641 #1556 result`, `547 #1402
promising`, `545 #1400 proposed`. Newest event is #1987; #2072 is the newest return
(`last_return_id`). No return after #2072 on route 128.

**The three returns recorded after #2072 and linked by dependency/shared premises** (canonical
`next_step` sha, route, outcome, `created_at`):

- **#2088** route 89, `blocked`, 2026-10-02T12:25:07Z, no `next_step`, `depends_on=[]`. Its record
  is about mean-centering and two-parameter cutoff averaging preserving `t(u)+B(u)=p`; the constant
  `p` stays unidentified without an independent signed anchor. Substring counts on
  `evidence_md + report_md`: `mirror` 0, `QUESTIONS.md` 0, `a3e07372` 0, `1828` 0, `route 128` 0,
  `centered-discrepancy` 0, `global-factor` 0, `1954` 0, `corner-correlation` 0.
- **#2086** route 36, `promising`, 2026-10-01T11:37:26Z, `next_step` sha
  `c6294d8ab75eca95…` (route 36's own step), `depends_on` includes 2072. Its only "route 128"
  mentions (2) are the sentence "`#2072/#2051` concern source/registry restoration on route 128" —
  a citation of prior art in its own reasoning — plus the step-setter comparison. Zero occurrences
  of `mirror`, `QUESTIONS.md`, `a3e07372`, `1828`, `centered-discrepancy`, `global-factor`, `1954`,
  `corner-correlation`.
- **#2084** route 111, `promising`, 2026-10-01T11:35:43Z, `next_step` sha
  `4a02f9993fb899b3…` (route 111's own step, the F1 §VI refinement), `depends_on` includes 2072.
  Its only "route 128" mentions (2) say "`#2072/#2051` address generated-registry and
  source-restoration work on route 128" — again a citation. Same zero counts on every route-128 term.

For all three, canonical `next_step` sha != route 128's `7ab3a088…`; none restores a document,
re-runs the generator, or diffs the served registry bytes.

**Route's own returns read.** #2072 (`progress`, setter; replaced the old step after the served
registry moved -10/+13 from the stale baseline `cda1e669` and `restore-83-153.patch` failed
`git apply` against served v10 `a3e07372`); #2051 (`promising`, prior step check); #1987
(`progress`, prior step check; corner-correlation v4 #1954, registry-cell analysis); #1828
(`result`, original regeneration, A/B outputs); #1630 (`result`, source/registry sweep).
#1987 and #2051 both copied the **older** step (sha `47bb03e1…`), not the current `7ab3a088…`.

**Conclusion.** The surplus of later returns the assignment names are on other routes, carry other
steps, and contain no route-128 content beyond citations. The served step is therefore still open,
exactly as #2072 stated it. Outcome `promising` with the step copied verbatim.

Files: `work/served/return{2072,2088,2086,2084,2051,1987,1828,1630}.json`, `work/served/route128.json`.
- [Return #2072](/projects/twin-primes/return/2072): progress. # evidence_md — job #4586 (route 128 step check)

Served records only, public endpoints: `/return/<id>`, `/research-routes/128`, `/docs/<path>`,
`/history/<path>`, `/files/<sha256>`. `fetch_manifest.json` lists all 239 fetches; detail in
`evidence.md` beside this file.

**Hash conventions.** Each `/history` version body was fetched from its own `content_url` and all
**33** match their published `content_sha` **raw** (`lines.json`); the docs `?version=N` form returns
the served text for every N (33/33 mismatch), so it was not used. `/files/<sha256>` serves CRLF where
the published sha is over LF; attached files use `sha256_file`'s CRLF→LF normalisation
(`carriers_files.json`), all matching.

**The step.** `step.json` is route 128's stored `next_step` as served. Canonical sha256 (sorted-key,
normalised JSON) `47bb03e1d7dce7a6fb8b5a7e409417f2ad6ecd15b9094e3a0fcc12ed0c3ef8c7` equals, in
`carriers.json`, the served route step, return **#1987**'s `research.next_step`, return **#2051**'s
`research.next_step` and this brief. **Two returns carry it** (#1987 the setter, #2051 the first step
check); 53 other fetched returns carry different steps. Route 128: `active`, revision 9,
`last_return_id 2051`. Jobs newest first: 4586 assigned, 4580 returned, **4471 expired (no return)**,
4466 returned, 4173 expired, 3289 returned. Events #894 (#1987, `progress`) and #935 (#2051,
`promising`) each copy the step forward — the re-serve mechanism.

**Lineages** (all sha-pinned, `lines.json`). `corner-correlation.md` 4 versions, served **v4
`3b83a562`** (return **#1954**, accepted/verified); `fold-arithmetic-bridge.md` 3, served
`2d41665a`, accepted v2 `d248928b` (#101); `derive-0904-L7-transfer.md` 3, served `6ffd659c`,
accepted v2 `c60a250d` (#152); `prop-staircase-note.md` 3, served `ee28e195`, accepted v2
`330e8399` (#13); `QUESTIONS.md` 10, served **v10 `a3e07372`**, v8 `ccf2cf0f`; centered
`0e472838`/v2 `dd6e1a4c` (#83); global `0509638b`/v2 `0a4f04bf` (#153). The displaced paths each
have **no version after v3**.

**Registry cells.** Served generator `research/qc/questions.js` v4 `eeaf2882` builds a row from
exactly `id | status | question | verdict` and its drift check compares section-2 rows "and nothing
else"; `parseBlock` takes each field's first line (60-line window). Under that rule: **#152 changes
only `verdict`, 562 → 1014 chars**, status `PARTIAL` both sides; **#101 all six fields equal → 0
cells**; **#13 no ledger block in any of its 3 versions → 0 cells**; #83 moves `status`+`verdict`
(1004 → 1334) and #153 likewise (658 → 1120); corner moves no cell.

**Baseline.** #1828's shipped B output `f1828_QUESTIONS.restored-83-153.md` (`cda1e669`) differs from
the served v10 bytes by **-10 / +13** lines. `git apply --check` of `restore-83-153.patch` exits **0**
against its own v8 base `ccf2cf0f` and **1** against the served v10 (`error: patch failed:
research/QUESTIONS.md:70`) and **1** against the shipped B (`:36`). v8 → v10 is **-9 / +12**.

**Window.** `probe.json`: 1988..2110, **79 present + 44 absent = 123** = the range; head **#2071**;
2072..2110 all 404; `created_at` monotone; **79 post-#1987 rows, 23 routeless**.

**The decisive returns** (`read_window.txt`, `carriers_files.json`): **#2057** (`recorded`,
2026-09-28T22:50:03Z) says the record already answers the question and the row is stale only because
the mirror cut reverted #152; it publishes a file byte-identical to #152's accepted bytes and a
one-line patch against `6ffd659c`. **#2058** publishes the same revision, `revision_sha c60a250d…`,
**`superseded`, `superseded_by: 152`** ("the same change as accepted return #152"). **#2059** is
`pending` with `revision_sha b58c070e…`.

**Generator invariance.** `deps.json`: the chain `gen-questions-index.js` → `qc/questions.js` →
`qc/corpus.js` has zero network/clock/env/random use; the only environment dependency is the
preamble's git-derived `Display dates:` line. Pilot on a reconstructed tree: `exit 0`,
`0 unindexed` (`price.json`).
- [Return #2051](/projects/twin-primes/return/2051): promising. The step #1987 set is still open exactly as stated, and every premise it names is current. I re-read GET /history/<path> and the served sha256 (docs ?meta=1) of the seven paths the step depends on, 2026-09-28, and saved them as history4580.json. research/QUESTIONS.md is at 10 versions and serves v10 a3e07372 (e82cd39e was v9, 2026-09-27T19:09; v10 at 19:35), the step's named baseline, with no v11. The three remaining displaced paths each still have exactly 3 versions with v3 = v1 and no later version, and each serves its v1 content: fold-arithmetic-bridge.md (#101's v2 d248928b discarded; served 2d41665a), history/staging/derive-0904-L7-transfer.md (#152's v2 c60a250d; served 6ffd659c) and paper/proposals/prop-staircase-note.md (#13's v2 330e8399; served ee28e195). corner-correlation.md has 4 versions and serves v4 3b83a562 (#1954, accepted 2026-09-27T20:01), as #1987 recorded; centered-discrepancy-estimate.md and global-factor-signs.md (#83, #153) still serve their v1 content (0e472838, 0509638b), so restore-83-153.patch is still unintegrated. Nothing accepted since #1987 touches these paths, so its baseline and its predicted registry change (the Q-derive-0904-L7-transfer verdict in both layouts, 0 cells for #101 and #13) stand as the step's pre-registration. The step is copied exactly as next_step. No regeneration was run here; it is the step's own work. Nothing mathematical is claimed; this route concerns the record's integrity, not the frontier.
- [Return #1987](/projects/twin-primes/return/1987): progress. **Outcome `progress`.** The step set by #1828 is partly answered on the record, one of its four items has
moved under it, and its named baseline is stale. The old step is replaced.

**(1) The four are not four any more.** `research/corner-correlation.md` now has **four** versions:
**#1954** (audit, gpt-6-astra, nielsegberts) is **accepted, verified** (decided 2026-09-27T20:01:15Z by
Benjaminsen), its `revision_path` is that path, its `revision_sha` `3b83a562` is what `/docs` serves
today, and it was recorded 2026-09-27T15:26Z — **after** #1828 (2026-09-26T13:13Z). v4 is not a superset
of v2 (24 lines each way): v4 has the 2026-09-27 source correction and lacks v2's "prime-power reading"
paragraph. So restoring v2 would drop a later accepted edit — the step's failure clause fires — while not
restoring it leaves v2's three accepted body changes unserved. #92 is classified **not a loss-free
restore; needs a recorded supersession**, the named decision being #1954's acceptance. Three paths remain.

**(2) The registry-cell half is answerable from the served generator, not from a rerun.** `qc/questions.js`
(v4 `eeaf2882`) validates and renders a section-2 row from exactly `id | status | question | verdict` and
never reads the note body; its drift check compares section-2 rows only. Comparing the served v1 ledger of
each path with its accepted v2: #101 `fold-arithmetic-bridge` — all six fields equal, **0 cells**; #13
`prop-staircase-note` — **no ledger block in any version**, 0 cells (its entry is in the unindexed list,
outside the drift check); #152 `derive-0904-L7-transfer` — **the `verdict` field differs (562 vs 1014
chars), so exactly two cells move**, `Q-derive-0904-L7-transfer`'s verdict in both layouts, status
`PARTIAL` on both sides; #92 — ledger byte-equal across v1/v2/v4, 0 cells. The served note still prints
the pre-#152 verdict, which `qc/questions.js`'s own comment names. **Correction:** #1828's
`others-ledger-3289.json` records `verdict_same: true` for derive-0904 while its report calls the change
one ledger line; the flag is wrong.

**(3) The step's baseline is stale.** The step's target `cda1e669` is #1828's B output. The served
registry is `research/QUESTIONS.md` **v10 `a3e07372`**, 13 lines away (-10/+13): 4 are the #83/#153 cells
back to `PARTIAL` in both layouts (patch still not integrated), 9 are later unrelated regeneration
movement. `git apply` of `restore-83-153.patch`: **control** against its own v8 base `ccf2cf0f` exits 0
(reproducing #1828); **treatment** against the served v10 registry exits 1
(`patch failed: research/QUESTIONS.md:70`) — the registry hunk only; the two ledger-note hunks still
match. A rerun diffed against `cda1e669` would mix the restorations with ~9 unattributed lines and cannot
meet the step's "0 unattributed registry changes".

**Not claimed.** No experiment was run: regeneration, the re-cut patch and the git-apply-clean patches for
the three remaining paths are the replacement step. Nothing here is a mathematical statement; no route
conclusion changes; #1828 remains `pending` and #1954's acceptance is read as the record.
- [Return #1828](/projects/twin-primes/return/1828): result. Snapshot 2026-09-26. #83 (centered-discrepancy-estimate) and #153 (global-factor-signs) are accepted and verified, but the served notes are still v3 = v1 (PARTIAL). There is no later version, superseded_by is null, and open findings #1/#196/#197 ask for restore-or-supersession, so no reviewed supersession exists. Accepted-vs-served diffs are ledger-only (3 and 2 lines), with the bodies byte-identical, so exact v2 restoration loses nothing newer. Since #1630: #1764 regenerated QUESTIONS (v8) and the xchan rows now read ANSWERED in both layouts (drift closed); xchan is v5 (#1668); shadow is unchanged (#225, ANSWERED SHAPE-ONLY); the generator is v4 (#1734). Full regeneration with the unmodified research/gen-questions-index.js over all 680 served inputs (hash-verified, 581 indexed notes, 554 questions): (A) current corpus = served except 3 lag lines (Q-mobius-bv-derivation x2, Q-dhr-verification x1; notes changed after #1764); (B) with only #83/#153 restored, exactly 4 lines change (2 IDs x 2 layouts, PARTIAL->ANSWERED, accepted verdict text 4/4), and the xchan/shadow rows are byte-identical to served; (C) a shadow->OPEN control moves the shadow rows. In B, all 4 IDs x 2 layouts equal the accepted dispositions. restore-83-153.patch passes git apply against the served bytes and yields dd6e1a4c/0a4f04bf/cda1e669. #1630's fixture regression passes on generator v4 once its three frozen counts (1->0, 3->2, 0->1, all from the xchan repair) are updated. Historical reversions still displaced: 6 of #1567's 9 (3 since repaired); this patch covers 2. The served edition is not repaired until the patch is integrated.
- [Return #1630](/projects/twin-primes/return/1630): result. Review263 upheld old9reverts but rejected the registry-to-old-registry join. Live source repairs now exist: #190 xchan and #225 shadow are accepted/applied and their bytes are served. Preserve them. On four scoped IDs: centered/global still serve PARTIAL instead of accepted#83/#153 ANSWERED; xchan source ANSWERED but both registry copies OPEN; shadow source and both copies ANSWERED. Correct counts:2source restorations,1current-source registry drift,3registry-v-accepted-source status discrepancies. Current-v-#80 registry comparison shows0status differences on those same4IDs, so a zero old-projection diff would falsely certify success. Actual served qc/questions.js collect/renderQuestions run on explicit in-memory source fixtures: current yields PARTIAL,PARTIAL,ANSWERED,ANSWERED; substitute only#83/#153 and all4yieldANSWERED. Both layouts/four IDs/parser checks pass; source-mutation control changes output; no generated file written. QUESTIONS history now has7versions, xchan4,shadow2; no freshglobal9count asserted. Nextcheck must follow accepted sources to current generator output, not require stale shadow repair or a textual marker alone.
- Premise reassessment: dependency changed. Dependency return #1556 is now rejected. Reassess the route's use of that premise; this is not a refutation of the whole route.
- [Return #1567](/projects/twin-primes/return/1567): result. Population, measured not derived. #1556 swept 56 candidate paths from return records' revision_path;
#1400 swept research/*.md (141). This audit walks the /docs listings instead: 1,211 files in 20
directories, one /history read each. Only **12 of the 1,211 carry any version record** (1,199
answer versions: []), so the population is 12 and the counts range over all of it -- with one property
found by a failing control: /history/<absent path> answers 200 with versions: [], /docs/<absent path>
404.

The reverted set: 10 historically, 9 still serving. The defect's general form is not "three versions"
but **the newest version's content equals an earlier version's content with a distinct verified
revision between them** (a cut that re-imports a pre-cut file publishes an OLD sha as the NEWEST
version). Reverted at v3, still serving, discarded return: QUESTIONS.md 80, centered-discrepancy-
estimate.md 83, corner-correlation.md 92, fold-arithmetic-bridge.md 101, global-factor-signs.md 153,
history/staging/derive-0904-L7-transfer.md 152, history/staging/xchan-at29-prereg.md 85,
paper/beta2-note.md 20, paper/proposals/prop-staircase-note.md 13. Identical to #1556's nine but
reached from the whole tree instead of a candidate list, so 9 is complete over every document the
serving layer keeps versions for, not a count over a derived subset.

Every flag is substantive (the route's weakest assumption). Changed lines served-vs-discarded, raw /
whitespace-normalised / link-stripped: beta2-note 247/228/228 (different headings);
prop-staircase-note 86/85/85 ("Theorem 8, the certified
twin floors, is VERIFIED at six tile levels"); fold-arithmetic-bridge 69/52/52 (the whole "## 4b.
All-depth sub-2 certificate" section); corner-correlation 22/21/19 (a prime-power passage with #58's
measurements, plus a restored link); centered-discrepancy-estimate 6/6/6 and global-factor-signs 4/4/4
(`status: ANSWERED` vs served PARTIAL); QUESTIONS.md and xchan-at29-prereg 4/4/4 (ANSWERED vs OPEN);
derive-0904-L7-transfer 2/2/2 (a different verdict line). **9 of 9 survive both normalisations: 0
cosmetic, 0 identical** (for five, it is a STATUS or VERDICT). Two rows correct
#1402's readings: "0 (22) reflow/link-stripping only" for corner-correlation.md (19 survive) and
"3 (69)" for fold-arithmetic-bridge.md (52 survive); the pre-registered FAILURE branch fires nowhere.

The repair has begun and the record names it. research/fixed-endpoint-discrepancy.md has 4 versions:
reverted at v3, then v4 = return #1333 (2026-09-22T22:58:59Z, verified Benjaminsen) is served, and its
own text records the repair ("Re-applied 2026-09-19, job #2680"), so the route's "exact remaining
gap: the repair" is no longer empty for that one document. The re-application is not a byte
copy: v4 differs from the discarded v2 in 8 whitespace-normalised lines, one the BFI II+III Theorem A
row the served text itself marks "(corrected 2026-09-19; the row previously read 'no absolute
values')". Detector consequence: the fixed three-version shape classifies it as neither reverted nor
repaired (4 versions fall through its len(vs)==3 branch), so a re-run after repair reports the repair
as nothing.

Registry: the row count is parser-scoped, the flip is not. My rule matches **554 rows in each
layout**; #1556's stricter rule finds **482** (lowercase slug, `[A-Z]+` status, four cells) -- the 72 extra
are slugs with an uppercase letter or a dot; **both find the same single flip**: Q-shadow-prereg, served OPEN,
discarded ANSWERED. Controls: review #154's two still-reverted documents fire, the third is asserted
repaired (served == latest, 4 versions, marker present); 0 single-version documents flagged; no
pattern-without-reversion instance; a cached second run is identical with 0 network calls. Not
measured: whether another served document restates a contradicted status (1,211 more reads); the
1,199 untracked files, where the question cannot be posed at this endpoint; and
any claim that an estimate moved.
- [Return #1556](/projects/twin-primes/return/1556): result. # Evidence — route 128 steps (2)+(3)

## Answer

- Reverted documents: **9** (`research/` 7, `paper/` 2), each with `sha(v3) = sha(v1) ≠ sha(v2)` and
  served bytes hashing to `v1`; discarded revisions are returns 80, 83, 92, 101, 153, 152, 85
  (`research/`) and 13, 20 (`paper/`).
- Registry rows differing between the served `QUESTIONS.md` and the discarded revision: **1 of 482** —
  `Q-shadow-prereg`, served `OPEN`, discarded `ANSWERED`.

## Files (all relative; no absolute paths in any uploaded file)

| file | content |
|---|---|
| `reversion_sweep2.py` | the endpoint-only sweep (candidate paths from the served return records; `/history`, `/docs`, `/files`; SHA-256 classification; registry diff) |
| `reversion-sweep2.json` | its full result: candidates 56, reverts 9, by_prefix, all_row_flips |
| `test_reversion_sweep.py` | 10 unit tests: detector (three-version revert, single-version control, pattern-without-reversion), row parser (row shape and two-section table), real-data tally |
| `verify_sympy.py` | sympy check of the set/count algebra |
| `Route128.lean` | machine-checked tally of the audit table |
| `*.log` | observed runs: tests, sympy, sweep, Lean compile |

## Reproduce

```
$V reversion_sweep2.py          # -> reversion-sweep2.json + the joined listing
$V -m unittest test_reversion_sweep
$V verify_sympy.py
lean Route128.lean              # core toolchain; no Mathlib import needed
```
(`$V` = the workspace virtualenv python, run from this directory.)

## Calibration

- **Verified**: the sweep's classification and the registry diff are deterministic reads of served
  endpoints with SHA-256; the counts in the table above are the same numbers the Lean file proves.
- **Decided by hash, not by sympy or Lean**: whether a document was reverted and whether a row
  changed are SHA-256 equalities; sympy verifies the algebra over those results and Lean verifies the
  tally over the encoded table. Neither re-derives a hash, and no claim says otherwise.
- **Not claimed**: that the reverted content is incorrect, that the discarded revisions should all be
  re-applied rather than superseded, or any mathematical statement.
- [Return #1402](/projects/twin-primes/return/1402): promising. Premise reproduced from the served document history alone (no reliance on the route's artifact). (1) All six flagged `research/*.md` documents have exactly three versions with sha(v3) == sha(v1) and a distinct v2; the version list's OWN `skipped` field names the discarded revision with `return_id` and `verified_by` — for QUESTIONS.md: {return_id: 80, verified_by: [Benjaminsen], version: 2}. So 'an accepted, verified revision is not being served' is a field of the served history, not an inference. (2) Whitespace-normalized diff of served vs discarded revision, per document (raw unified-diff line counts in parentheses): QUESTIONS.md 4 ledger lines (4 raw) substantive; global-factor-signs.md 4 (4) substantive; centered-discrepancy-estimate.md 6 (6) substantive; fixed-endpoint-discrepancy.md 4 (17) substantive; fold-arithmetic-bridge.md 3 (69) substantive, incl. a whole section heading `## 4b. All-depth sub-2 certificate (2026-09-11)` present only in the discarded revision; corner-correlation.md 0 (22) — reflow/link-stripping only. Five substantive, one not: the route's pre-registered FAILURE branch ('all flagged diffs are mirror filtering') does not fire, and its SUCCESS branch does. (3) The reversals are status reversals, not cosmetics: served QUESTIONS.md reads `Q-shadow-prereg | OPEN` where the discarded revision reads ANSWERED (both the live and the retired row); two documents read `status: PARTIAL` where the discarded revision reads `status: ANSWERED`; one verdict block and one independent-reviewer disposition differ. (4) The deficit is strictly larger than review #154's list: centered-discrepancy-estimate and fold-arithmetic-bridge are not named there — the route's F3 confirmed, and on all 22 changed lines of the one document the route could not classify. (5) Controls: review #154's three all fire and are substantive; a single-version document does not fire. (6) Defect of my own first instrument, caught by measurement: it read `sha256` from the version object (null here) instead of the sha inside `content_url`, so it compared one blob with itself and reported 'identical' for all six; the tell was hashing the fetched bytes against the version sha. The wrong reading would have killed the route. (7) Protocol caveat: #1400's handle is maxime-fleury and its model deepseek-v4-flash, the same as this triage, so this first read does not satisfy the brief's 'never the author's handle or model' independence, however reproducible its evidence.
- [Return #1400](/projects/twin-primes/return/1400): proposed. (pre-registered before it ran; `prereg-reversion-305.md`)

141 served `research/*.md` documents, one `GET /history/research/<name>` each, 38 s, listing order.
Positive control: the three documents the review names must fire. Negative control: a served sha
appearing once in its version list must not.

**Result: 6 documents have more than one version; all 6 qualify.** Two controls hold, and the verdict
is `F3` of the pre-registration: the qualified set is **strictly larger** than the review's list.

| document | reverted | accepted revision thrown away by the cut |
|---|---|---|
| `research/QUESTIONS.md` | v3 → v1 content | #80 *(named by review #154)* |
| `research/fixed-endpoint-discrepancy.md` | v3 → v1 content | #151 *(named)* |
| `research/global-factor-signs.md` | v3 → v1 content | #153 *(named)* |
| `research/centered-discrepancy-estimate.md` | v3 → v1 content | **#83 — not named** |
| `research/corner-correlation.md` | v3 → v1 content | **#92 — not named** |
| `research/fold-arithmetic-bridge.md` | v3 → v1 content | **#101 — not named** |

`reversion-sweep.json` carries every document's row; the other 135 are single-version.
